Soils and Foundations 2013;53(4):607–615

The Japanese Geotechnical Society

Soils and Foundations

www.sciencedirect.com
journal homepage: www.elsevier.com/locate/sandf

Pipe pile setup: Database and prediction model using artificial

neural network
Bashar Tarawnehn
Civil Engineering Department, The University of Jordan, Amman 11942, Jordan
Received 31 August 2012; received in revised form 14 April 2013; accepted 22 May 2013
Available online 25 July 2013

Abstract

Over the last few years, artificial neural networks (ANNs) have been applied to many geotechnical engineering problems with some degree of
success. With respect to the design of pile foundations, the ability to accurately predict pile setup may lead to more economical pile design,
resulting in a reduction in pile length, pile section, and size of driving equipment. In this paper, an ANN model was developed for predicting pipe
pile setup using 104 data points, obtained from the published literature and the author's own files. In addition, the paper discusses the choice of
input and internal network parameters which were examined to obtain the optimum ANN model.
Finally, the paper compares the predictions obtained by the ANN with those given by a number of empirical formulas. It is demonstrated that
the ANN model satisfactorily predicts the measured pipe pile setup and significantly outperforms the examined empirical formulas.
& 2013 The Japanese Geotechnical Society. Production and hosting by Elsevier B.V. All rights reserved.

Keywords: Pile foundation; Pile setup; Artificial neural networks

1. Introduction in pile capacity regularly occurs. This phenomenon is referred to

Piles are relatively long and generally slender structural founda-
tion members that transmit superstructure loads to deep soil layers.
Piles usually serve as foundations when soil conditions are not
suitable for the use of shallow foundations. Moreover, piles have
other applications in deep excavations and in slope stability. The
ultimate axial load carrying capacity of the pile is composed of the
end-bearing capacity of the pile and the shaft friction capacity.
Driving a pile into the ground leads to disturbance and
displacement of the soil surrounding the pile. As the soil
recovers from the driving disturbance, a time-dependent increase
n
Tel.: +962 79 754 9696.
E-mail address: btarawneh@ju.edu.jo
Peer review under responsibility of The Japanese Geotechnical Society.
as “setup” by geotechnical engineers. A significant increase in
pile capacity could occur due to this phenomenon. Considering
pile setup in the axial load capacity of driven pile may lead to
more economical pile design, leading to reduction in pile length,
pile section, and size of driving equipments. Two mechanisms
have been claimed to explain the setup phenomenon:
1. The dissipation of excess pore water pressure for some
period of time after driving the piles. This is generated due
to soil remolding/disturbance during pile driving. The
associated increase in lateral effective stress with increasing
time gives rise to an increase in shear strength and thus the
axial capacity of the pile. Obviously the duration of this
reconsolidation depends on the permeability of the soil. It
ranges from days in coarse-grained soils to months or years
in fine-grained soils (Yan and Yuen, 2010).
2. The effect of soil aging: Aging refers to a time-dependent
change in soil properties at a constant effective stress. Chow

0038-0806 & 2013 The Japanese Geotechnical Society. Production and hosting by Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.sandf.2013.06.011
608 B. Tarawneh / Soils and Foundations 53 (2013) 607–615
et al. (1998) showed imperceptible increases in density
during the aging period, so the changes in behavior must be
attributable to micro-structural rearrangements of the sand
grains and their contacts with time. The changes in soil
properties due to aging are unlikely to be the major factor
controlling setup in sand.
Chow et al. (1998) suggested that the best explanation for
the marked setup effects on driven piles in non-cohesive soils
is that sand creep rather than climatic or tide-related changes in
pore pressure weakens the arching mechanisms surrounding
the pile shaft, increasing horizontal stresses while also produ-
cing stronger dilation effects during loading.
Axelsson (1998) showed that pile driving in sand can
generate strong arching effects, even at significant depths,
and then the arching deteriorates with time due to stress
relaxation and leads to an increase in horizontal stress on the
shaft. The increase in horizontal stress due to stress relaxation
can continue for several months and is approximately linear
with the logarithm of time. The other main reason is that soil
aging phenomenon with respect to piling causes the reorienta-
tion of particles, leading to interlocking. In other words, both
soil particles interlocking with the surface roughness and stress
relaxation provide an explanation of the strong setup effects of
pile in non-cohesive soils.
Ng et al. (2013a, 2013b) studied the pile setup through a
field investigation on five fully instrumented steel H-piles
embedded in cohesive soils. Based on the field data collected,
it was concluded that the skin friction component, not the end
bearing, contributes predominantly to the setup, which can be
estimated for practical purposes using soil properties, such as
coefficient of consolidation, undrained shear strength, and the
standard penetration test N-value. A new approach was
developed for estimating pile setup using dynamic measure-
ments and analyses in combination with measured soil
properties.
To characterize setup, geotechnical engineers are utilizing
dynamic monitoring using a Pile Driving Analyzer (PDA)
during initial driving and additional dynamic monitoring
during restrike testing several hours to a few weeks after
initial driving. On projects involving a large number of driven
piles, the savings in pile costs can significantly exceed the cost
of testing required to characterize setup. However, on projects
involving a small number of piles, the cost of testing to
characterize setup may exceed pile installation cost savings.
Therefore, on small projects, testing required to characterize
setup directly cannot be justified from an economic standpoint.
While numerous research projects are still being carried out
to study the underlying mechanisms causing the pile setup,
simple empirical relations are available in the literature to
predict this increase in pile capacity with time (e.g. Skov and
Denver, 1988; Svinkin and Skov, 2000). The relations predict
the pile capacity from the initial capacity (described as end of
driving, EOD) and elapsed time after driving, in which two
sets of model constants were suggested for clayey and sandy
soil, respectively, based on limited data sets. The reliability of
adopting these model constants is questionable. Also, the form
of existing empirical relations has certain weaknesses.
In this paper, Artificial Neural Networks (ANNs) are used to
predict pipe piles setup. The aims of the paper are to:

Compile pipe pile setup data from the author's files and the
published literature.

Develop an (ANN) model that can predict the pipe
pile setup.

Study the effect of ANN geometry and some internal
parameters on the performance of ANN models.

Explore the relative importance of the factors affecting
pipe pile setup prediction by carrying out sensitivity
analysis.

Compare the performance of the developed ANN
model with four of the most commonly used traditional
methods.
2. Estimation of pile setup
Many researchers have observed the increase of pile
capacity with time after pile installation into the ground. From
their field studies, they have developed empirical relationships
for predicting pile capacity with time if pile capacity at end of
drive is given. Below is a summary of the most commonly
used formulas.
2.1. Skov and Denver (1988)
Skov and Denver (1988) presented a formula that is a linear
relationship with respect to the log of time.
Qt ¼ Qo ½A logðt=t o Þ þ 1 ð1Þ
where
Qt ¼ axial capacity at time “t” after driving,
Qo ¼ axial capacity at time to,
A ¼ a constant depending on soil type, and
to ¼ an empirical value measured in days.
In the above equation, to is a function of the soil type and
pile size and is the time at which the rate of excess pore-water
pressure dissipation becomes uniform (linear with respect to
the log of time). The value of to is defined as 0.5 for sand and
1.0 for clay. And the value of parameter “A” is a function of
soil type, pile material, type, size, and capacity. The “A” value
is presented by 0.2 for sand and 0.6 for clay.
2.2. Svinkin (1996)
Svinkin (1996) developed a formula for pile setup based on
load test data:
Qt ¼ 1:4QEOD t 0:1 upper bound ð2Þ
 B. Tarawneh / Soils and Foundations 53 (2013) 607–615 609

Qt ¼ 1:025QEOD t 0:1 lower bound ð3Þ

2.3. Long et al. (1999)

Long et al. (1999) presented a formula considering to ¼ 0.01

days, which is modified from Eq. (1):
Qt ¼ QO ½A logðt=0:01Þ þ 1 ð4Þ

2.4. Svinkin and Skov (2000)

Svinkin and Skov (2000) proposed a formula for pile setup

based on Eq. (1):
Qt =QEOD −1 ¼ B½log10 ðtÞ þ 1 ð5Þ Fig. 1. Determination of the average vertical effective stress for the entire pile
length (Das, 2011).
The factor “B” is similar to the factor “A” in Eq. (1).
The factor “B” ranges from 1.6 to 3.5.

3. Pipe pile setup database
A database was compiled from the results of 104 pipe pile
dynamic tests and Case Pile Wave Analysis Program (CAP-
WAP) analyses for pile capacity. The dynamic testing consists
of instrumenting the pile during driving with accelerometers
and strain transducers, which are connected to a field-portable
digital microcomputer which processes the acceleration and
strain signals. The “raw” data as collected in the field is
capable of predicting combined shaft and toe resistance.
Additional laboratory analysis of field-measured dynamic
monitoring data called a CAPWAP analysis is capable of
predicting shaft resistance distribution, and toe resistance.
Since it collects data during driving, dynamic monitoring is
uniquely suited to determining capacity instantaneously at the
end of driving. Restrike testing can be provided for multiple
tests at various times after driving. When CAPWAP analyses
are performed on end-of-drive and restrike data, the distribu-
tion of setup along the shaft can be determined.
The data were obtained from the author's own files and the
published literature. Twenty pipe pile setup data were collected
from different projects in Ohio (Khan and Decapite, 2011) and
84 were collected from the published literature. The references
used to compile the database are given in Table 1.
The collected data included pipe pile diameter, driven
length, time after installation (t), soil type, the average vertical
effective stress for the entire pile embedment depth, the initial
Table 1
Database references.
Reference Location of tests No. of testing
Antorena and McDaniel (1995) Florida 4
Thompson et al. (2009) Mississippi 58
Author's own files Ohio 20
(Khan and Decapite, 2011)
Dover and Howard (2002) California 11
Holloway and Beddard (1995) California 5
Komurka (2004) Wisconsin 6
axial capacity (described as end of driving, QEOD) the axial
capacity at time “t” after driving (Qt), and the capacity increase
due to setup.
The soil was divided into two major groups based on the
AASHTO classification: Granular materials (A-1, A-2, and
A-3) soils and Silt-clay materials (A-4, A-5, A-6, and A-7)
soils. The values associated with the soil type are “1” for
Granular materials, “ 2” for Silt-clay materials, and others are
calculated by considering percentage of soil type (Granular,
Silt-clay) up to the pile tip.”
The average vertical effective stress for the entire pile
embedment depth can be determined using the following
equation and Fig. 1:
Average vertical effective stress ¼ ðA1 þ A2 þ A3 Þ=L ð6Þ
4. Artificial neural networks
Over the past two decades, artificial neural networks
(ANNs) have emerged as powerful and versatile computational
tools for organizing and correlating information in ways that
have proved useful for solving certain types of problems that
are difficult to tackle using traditional numerical and statistical
methods (Transportation Research Board, 1999).
ANNs consist of a group of artificial neurons that are
interconnected in a way similar to the architecture of the
human brain. This computational technique has the ability to
recognize, capture and map features known as patterns
contained in a set of data, mainly due to the high interconnec-
tions of neurons that process information in parallel (Zeghal
and Khogali, 2005). A network that has learned the patterns
defining the relationship between the input and output of a
certain test or process can later be used to predict new
conditions for which the output is not known. An ANN
system consists of three or more layers. The first layer has
the input neurons (parameters), while the last layer contains the
output. In addition, one or more layers can be between the
610 B. Tarawneh / Soils and Foundations 53 (2013) 607–615
input and output layers, which are known as the hidden layers.
Those layers form the network's means of delineating and
learning the patterns governing the data that the network is
presented with.
There are many ways a neural network can be trained. The
back-propagation technique, which was developed by
Rumelhart et al. (1986), is the most popular process and has
been used in many fields of science and engineering. With this
method, the weights of the network are adjusted during the
training phase to minimize error. In each iteration, the error
propagates backward to minimize the error to a desired level.
In recent years, there has been a growing interest in the use
of ANNs in the design of deep and shallow foundations. Ornek
et al. (2012) presented a study to describe the use of artificial
neural networks (ANNs), and the multi-linear regression model
(MLR) to predict the bearing capacity of circular shallow
footings supported by layers of compacted granular fill over
natural clay soil. Nejad and Jaksa (2011) developed an ANN
model to predict pile settlement based on standard penetration
test data. Abu-Kiefa (1998) introduced three neural network
models to predict the capacity of driven piles in cohesionless
soils. Lee and Lee (1996) utilized neural networks to predict
the ultimate bearing capacity of piles. Chan and Chow (1995)
developed a neural network as an alternative to pile driving
formula. Goh (1994a, 1995b) presented a neural network
model to predict the friction capacity of piles in clays.
5. Development of ANN model to predict pipe pile setup
The development of an ANN model requires the determina-
tion of model inputs and outputs, division and pre-processing
of the available data, the determination of appropriate network
architecture, stopping, and model validations. Neuro-Solutions
6.0 Software was used in creating the neural network models.
This software combines a modular design interface with
advanced learning procedures, giving the power and flexibility
needed to design the neural network that produces the best
solution.
The data used to develop and validate the neural network
model were obtained from both published literature and the
author's own files. Suitable case studies were those having pile
load tests and information regarding the piles and soil. The
database contains a total of 104 cases of pile load tests. The
references used to compile the database are given in Table 1.
5.1. Model inputs and outputs
In order to achieve accurate pile setup predictions, a
thorough understanding of the factors affecting pile setup is
needed. Most of traditional pile setup methods include the
following fundamental parameters: pile geometry, pile mate-
rial, average effective stress at tip, soil properties, time, and
QEOD.
Setup is much affected by soil and pile type. Effective stress
changing with time also has a significant role in the evaluation
of pile capacity. The pile length and pile diameter affect setup
influence zone (Jongkoo, 2007).
In this paper, the factors that are presented to the ANN
as model input variables are pile diameter, pile length, soil
type, average effective stress at tip, and the time. Pile capacity
increase due to setup (Qt−QEOD) was the single output
variable.
To identify which of the input variables have the most
significant impact on pile setup prediction, a sensitivity
analysis was carried out on the trained network. A simple
and innovative technique proposed by Garson (1991) was used
to define the relative importance of the input variables by
examining the connection weights of the trained network. The
results of the sensitivity analysis are discussed later.
5.2. Data division and pre-processing
The data were divided into three sets; training, cross
validation, and testing. Seventy percent of the data points
were selected for training, 15% were selected for cross
validation, and 15% were used for testing the network. The
training data points were used to train the network and
compute the weights of the inputs. The test data points were
used to measure the performance of the selected ANN model.
The cross validation computes the error of an independent data
set (cross validation data set) at the same time that the network
is being trained with the training set.
During training, the input and desired data will be repeatedly
presented to the network. As the network learns, the error will
drop towards zero. Lower error, however, does not always
mean a better network. It is possible to over train the network.
The network can be over-trained to the point where perfor-
mance on new data actually deteriorates. Overtraining results
in a network that memorizes the individual exemplars rather
than trends in the data set as a whole. A stopping criteria needs
to be established to avoid overtraining. In this paper, cross
validation was used as the stopping criteria. To stop the
training of the network; part of the data was set aside for the
purpose of monitoring the training process and to guard against
overtraining. The cross validation set was never used for
network weight adjustment.
It is important that the data used for training, cross
validation, and testing represent the same population
(Masters, 1993) and the statistical properties (e.g. mean,
standard deviation and range) of the data subsets need to be
similar (Shahin et al., 2004). Also, ANNs perform best when
they do not extrapolate beyond the range of their training data
(Flood and Kartam, 1994; Tokar and Johnson, 1999). Accord-
ingly, in order to develop the best possible model, all patterns
that are contained in the data need to be included in the
training set. Similarly, since the test set is used to determine
when to stop training, it needs to be representative of the
training set and should contain all of the patterns that are
present in the available data (Shahin et al., 2002). To
accomplish this, several random combinations of the training,
cross validation and testing sets were tried until a statistically
consistent data set was obtained. The statistical parameters
considered include the mean, standard deviation, minimum,
maximum and range, as suggested by Shahin et al. (2004).
 B. Tarawneh / Soils and Foundations 53 (2013) 607–615 611

Even though numerous random combinations of training, cross
validation and testing sets were tried, there are still some slight
inconsistencies in the statistical parameters for each subset, as
shown in Table 2. Generally, the statistics are in good
agreement and the data set can be used to represent the same
population.
5.3. Selection of model architecture
Selecting the network architecture is one of the most
important and difficult tasks in ANN model development
(Maier and Dandy, 2000). It requires the selection of the
optimum number of hidden layers and the number of nodes in
each of these. There is no unified theory for determination of
an optimal ANN architecture (Shahin, 2003). The number of
nodes in the input and output layers are restricted by the
number of model inputs and outputs. A total of five input
variables are included in developing the ANN model, namely,
pile diameter, pile length, soil type, time and the average
effective stress. The output layer has a single node representing
the measured value of pile capacity increase due to setup.
Several network structures with different numbers of hidden
layers and nodes in the hidden layer were trained and tested.
This strategy was chosen to find the best performing network
architecture among different models. Although it has been
shown that a network with one hidden layer can approximate
any continuous function (Hornik et al., 1989) in this research
one and two hidden layers were employed. In order to
determine the optimum network geometry, first ANNs with a

Table 2
ANN input and output statistics.

Model variable and data sets Mean Standard deviation Min. Max. Range

Pipe pile diameter (cm) Training set 60.72 17.5 30.5 107 76.5
Cross Validation set 55.05 15.45 30.5 76.2 45.7
Testing set 60.76 18.57 30.5 107 76.5
Driven length (m) Training set 25.49 7.06 4.6 41.2 36.6
Cross Validation set 25.36 9.18 5.7 40 34.3
Testing set 24.76 10.26 8 41.6 33.6
Soil type Training set 1.64 0.39 1 2 1
Cross validation set 1.58 0.39 1 2 1
Testing set 1.47 0.46 1 2 1
Average vertical effective stress (kN/m2) Training set 231.18 62.85 41.3 383.62 342.32
Cross validation set 225.82 82.24 51.48 333.25 281.77
Testing set 224.61 90.68 71.75 336.11 264.36
Time (days) Training set 24.61 31.79 0.06 85 84.96
Cross validation set 14.28 18.34 0.5 70 69.5
Testing set 11.54 18.97 0.8 69 68.2
Capacity increase due to setup (Qt−QEOD) Training set 1927.12 1277.03 71.17 5150.6 5079.43
Cross validation set 2120.86 1548.61 275.79 4982 4706.21
Testing set 1795.14 1527.41 124.55 4919 4794.45

Table 3
ANN models with a single hidden layer.

Model No. Input nodes Hidden layer-1 Output layer Testing data

Processing elements Transfer function Processing elements Transfer function RMSE R

1 5 1 tanh 1 tanh 962.91 0.77

2 5 2 tanh 1 tanh 670.75 0.9
3 5 3 tanh 1 tanh 787.34 0.89
4 5 4 tanh 1 tanh 548.72 0.96
5 5 5 tanh 1 tanh 1099.78 0.76
6 5 6 tanh 1 tanh 858.60 0.85
7 5 7 tanh 1 tanh 1204.58 0.68
8 5 10 tanh 1 tanh 1647.89 0.49
9 5 1 Sig. 1 Sig. 835.47 0.83
10 5 2 Sig. 1 Sig. 731.82 0.88
11 5 3 Sig. 1 Sig. 1211.43 0.69
12 5 4 Sig. 1 Sig. 708.51 0.9
13 5 5 Sig. 1 Sig. 845.96 0.84
612 B. Tarawneh / Soils and Foundations 53 (2013) 607–615

0.88
0.89
0.89

0.87
0.88
0.86
0.92
0.91
0.88
0.87
single hidden layer and different number of nodes in the

0.9
0.9
R
hidden layer are trained with sigmoidal (Sig.) and hyperbolic

Testing data
tangent (tanh) transfer functions for the hidden and output

740.81
784.71
685.07
738.54
690.32
747.21
737.56
885.02
721.59
783.33
737.92
772.83
RMSE
layers. Combinations of numbers of elements in the hidden
layer and types of transfer function that yielded the most
accurate predictions of pile setup are shown in Table 3.
Then ANNs with two hidden layers with different numbers

Transfer function
of nodes in the hidden layers are trained. Models have been
trained with sigmoidal (Sig.) and hyperbolic tangent (tanh)
transfer functions for the hidden and output layers. Combina-
tions of the number of elements in the hidden layers and type

tanh
tanh

tanh

tanh
tanh
tanh

tanh
Sig.

Sig.

Sig.
Sig.
Sig.
of transfer functions that yielded the most accurate predictions
of pile setup are shown in Table 4.

Processing elements
5.4. Model optimization

Output layer
Training is the process of optimizing the connection
weights. The goal is to find a global solution to a non-linear
optimization problem. The method most commonly used for

1
1
1
1
1
1
1
1
1
1
1
1
establishing the optimum weight combination of feed-forward
neural networks is the back-propagation algorithm (Rumelhart

Transfer function
et al., 1986) which is based on first-order gradient descent.
Feed-forward networks trained with the back-propagation
algorithm have been applied successfully for many geotechni-
cal engineering problems (e.g. Goh, 1994; Najjar and Basheer,
tanh

tanh
tanh

tanh
tanh
tanh
tanh
tanh
Sig.

Sig.
Sig.
Sig.
1996). Accordingly, the back-propagation algorithm is used for
optimizing the connection weights in this study. Details of the
back-propagation algorithm can be found in many publications
Processing elements

(e.g. Fausett, 1994).

Hidden layer-2

The Levenberg–Marquardt (LM) algorithm was used as a

learning rule in this study. It is one of the most appropriate
higher-order adaptive algorithms known for minimizing the

10

12
6
6
6
4
4
6
4
2
8

4
mean square error (MSE) of a neural network (Principe et al.,
1999).
Transfer function

5.5. Stopping criteria

Stopping criteria are used to decide when to stop the training

process (Shahin et al., 2002) they determine whether the model
tanh
tanh

tanh

tanh
tanh
tanh

tanh
Sig.
Sig.
Sig.

Sig.

Sig.

has been optimally or sub-optimally trained (Maier and Dandy,

2000). Many approaches can be used to determine when to
stop training, such as those described by Shahin et al. (2002).
Processing elements

As mentioned earlier, the cross-validation technique (Stone,

Hidden layer-1

1974) is used in this study, as it is considered that sufficient

data are available to create training, testing and validation sets
and it is the most valuable tool to ensure over-fitting does not
10

12
6
6
6
4
4
9
6
4
8

occur (Smith, 1993). The training set is used to adjust the

ANN models with two hidden layers.

connection weights, whereas the testing set measures the

ability of the model to generalize. Using this set, the
Input nodes

performance of the model is checked at many stages during

the training process, and training is stopped when the error of
the testing set starts to increase (Shahin et al., 2002).
5
5
5
5
5
5
5
5
5
5
5
5

5.6. Model validation

Model No.
Table 4

Once training of the model has been successfully accom-

14
15
16
17
18
19
20
21
22
23
24
25

plished, the performance of the trained model should be validated

 B. Tarawneh / Soils and Foundations 53 (2013) 607–615 613

using data sets that have not been used as part of the learning
process. This data set is known as the testing set. The purpose of
the model validation phase is to ensure that the model has the
ability to generalize within the limits set by the training data in a
robust fashion, rather than simply having memorized the input–
output relationships that are contained in the training data (Shahin
et al., 2002).The coefficient of correlation, R; the root mean
squared error, RMSE; and the mean absolute error, MAE, are the
main criteria that are used to evaluate the prediction performance
of ANN models. The coefficient of correlation is a measure that is
used to determine the relative correlation between the predicted
and measured data.
The root mean square error, RMSE, is the most popular
measure of error. It has the advantage that large errors receive
much greater attention than small errors (Hecht-Nielson,
1990). The root mean square error, RMSE, and mean absolute
error, MAE, are desirable when the data evaluated are smooth
or continuous (Twomey and Smith, 1997).
6. Results
For the one hidden layer models, Table 3 shows the results
of the top performing models. The results showed that the
Table 5
Sensitivity analysis of the relative importance (%) for the ANN optimal model.
ANN input variable Relative importance %
Pipe pile diameter 13
Driven length 38
Soil type 12
Average vertical effective stress 23
Time 14
RMSE values were between 548.72 and 1647.89 and the
coefficient of correlation, R, values were between 0.49 and
0.96 for the testing data set. It is noted that model 4, which
used four processing elements in the hidden layer and tanh as
transfer function for both hidden and output layer was the best
performing among all models. It has the lowest RMSE value
(548.72) and the highest R value (0.96) for the testing data set.
The RMSE for the training and cross-validation data sets were
0.22 and 0.57, respectively, for model 4.
Table 4 shows the results of the top performing ANN
models with two hidden layers. The results showed that the
RMSE values were between 685.07 and 885.02 and the
coefficient of correlation, R, values were between 0.86 and
0.92 for the testing data set.
It is noted that model 16, which used six processing
elements in each hidden layer, sigmoid as transfer function
for the first hidden and output layers, and tanh as transfer
function for the second hidden layer, has the lowest RMSE
(685.07); R value for this model was 0.89 for the testing
data set.
Model 23, which used 10 processing elements in each
hidden layer, sigmoid as transfer function for the first hidden
Table 6
Coefficient of determination for measured Qt vs. predicted.
Prediction method Coefficient of determination, R2
ANN-Model-4 0.92
Skov and Denver 0.8
Long 0.82
Svinkin-upper 0.66
Svinkin-lower 0.66
Svinkin and Skov 0.81

9000

8000

7000

6000
Measured Qt (kN)

5000

4000

3000

2000

1000

0
0 1000 2000 3000 4000 5000 6000 7000 8000 9000
Predicted Qt (kN)
ANNs Optimal Model Skov and Denver Long
Svinkin-Upper Svinkin-Lower Svinkin and Skov

Fig. 2. Measured versus predicted Qt for the optimal ANN model; Skov and Denver (1988), Svinkin (1996), Svinkin and Skov (2000), and Long et al. (1999)
formulas.
614 B. Tarawneh / Soils and Foundations 53 (2013) 607–615
and output layers, and tanh as transfer function for the second
hidden layer, has the highest R value (0.91); the RMSE value
for this model was 783.33 for the testing data set. The RMSEs
for the training and cross-validation data sets were less than
0.25 and 0.6, respectively, for models 16 and 23.
It can be concluded that model 4 is the best performing
optimal model among all ANN models. Based on the available
data and result, this model can be recommended to predict the
pipe pile capacity increase due to setup.
The results of a sensitivity analysis for the optimal model
(model 4) with a single hidden layer and four nodes are shown
in Table 5. As one would expect, it can be seen that the pile
length, average vertical effective stress, time, pipe pile
diameter, and the soil type have the most significant effect
on the predicted pipe pile capacity increase due to setup.
7. Numerical model
In this section, the testing data set was used to compare the
measured Qt with the predicted Qt using the optimal ANN
model (i.e., model 4, with one hidden layer and four nodes in
the hidden layer), as well as Qt predicted using the empirical
formulas presented in Section 2 of this paper. It is known that
this data set was not included in the training data of the network.
Plots of measured versus predicted Qt are shown in Fig. 2 for
the optimal ANN model and the Skov and Denver (1988),
Svinkin (1996), Svinkin and Skov (2000), and Long et al.
(1999) formulas. It can be seen that the optimal ANN model
satisfactorily predicted the measured data and significantly
outperformed the examined empirical formulas.
Table 6 shows the coefficient of determination, R2 for the
measured Qt versus the optimal ANN model and the Skov and
Denver, Svinkin, Svinkin and Skov, and Long et al. formulas. It
can be seen that the optimal ANN model has the highest
R2 value.
8. Conclusions
A back-propagation neural network was used to examine the
feasibility of ANNs to predict the pipe pile axial capacity
increase due to setup. A database containing 104 case records
of field measurements of pipe pile setup was used to develop
and verify the model. The results indicated that back-
propagation neural networks have the ability to predict the
settlement of pile with an acceptable degree of accuracy
(R ¼ 0.96, RMSE ¼ 548.72).
The ANN method can be used as a very good tool for
estimating pile setup. The ANN Model adequately predicted
pile setup and significantly outperformed the examined empiri-
cal formulas. The result of this study indicated that ANNs
yielded more accurate pile setup predictions than those
obtained from the traditional methods examined: Skov and
Denver (1988), Svinkin (1996), Svinkin and Skov (2000), and
Long et al.'s (1999) formulas.
A sensitivity analysis indicated that, as one would expect,
pile diameter, pile length, soil type, average effective stress at
tip, and time have the most significant effects on the pipe
pile setup.
References
Abu-Kiefa, M.A., 1998. General regression neural networks for driven piles in
cohesionless soils. Journal of Geotechnical and Geoenvironmental Engi-
neering – ASCE 124 (12), 1177–1185.
Antorena, J.M., McDaniel T.G., 1995. Dynamic pile testing in soils exhibiting
set-up. In: Deep Foundations Institute 20th Annual Members Conference
and Meeting. Charleston, South Carolina, pp. 17–27.
Axelsson, G. 1998. Long-term increase in shaft capacity of driven piles in
sand. In: The 4th International Conference on Case Histories in Geotech-
nical Engineering, St. Louis, Missouri, pp. 301–308.
Chan, W., Chow, Y.K., 1995. Neural network: an alternative to pile driving
formulas. Journal of Computer and Geotechnics 17 (2), 135–156.
Chow, F.C., Jardine, R.J., Brucy, F., 1998. Effects of time on capacity of pipe
piles in dense marine sand. Journal of Geotechnical and Geoenvironmental
Engineering – ASCE 124 (3), 254–264.
Das, B.M., 2011. Principles of Foundation Engineering, Stamford, CT, USA.
Dover, A., Howard, R. 2002. High Capacity Pipe Piles at Sanfrancisco
International Airport. Deep Foundations Congress, Geotechnical Special
Publication, ASCE: 655–667.
Fausett, L., 1994. Fundamentals of Neural Networks: Architecture, Algo-
rithms, and Applications. Prentice-Hall, Englewood Cliffs, NJ.
Flood, I., Kartam, N., 1994. Neural networks in civil engineering: principles
and understanding. Journal of Computing in Civil Engineering – ASCE 8
(2), 131–148.
Garson, G., 1991. Interpreting neural-network connection weights. AI Expert 6
(7), 47–51.
Goh, A., 1994. Seismic liquefaction potential assessed by neural networks.
Journal of Geotechnical and Geoenvironmental Engineering-ASCE 120
(9), 1467–1480.
Goh, A.T.C., 1994a. Nonlinear modeling in geotechnical engineering using
neural networks. Australian Civil Engineering Transactions 36 (4),
293–297.
Goh, A.T.C., 1995b. Empirical design in geotechnics using neural networks.
Geotechnique 45 (4), 709–714.
Hecht-Nielson, R., 1990. Neurocomputing. Addison-Wesley, Reading, MA.
Holloway, D.M., Beddard, D.L. 1995. Dynamic testing results, indicator pile
test program, I-880, Oakland, California. In: Deep Foundations Institute
20th Annual Members Conference and Meeting. Charleston, South
Carolina, pp. 105–126.
Hornik, K., Stinchcombe, M., White, H., 1989. Multilayer feed forward
networks are universal approximators. Neural Network 2, 359–366.
Jongkoo, J. 2007. Fuzzy and Neural Network Models for Analyses of Piles,
North Carolina State University.
Khan, L.I., Decapite, K. 2011. Prediction of Pile Set-up for Ohio Soils. Report
No. FHWA/OH-2011/3, FHWA, U.S. Department of Transportation.
Komurka, V.E., 2004. Incorporating set-up and support cost distributions into
driven pile design. Current Practices and Future Trends in Deep Founda-
tions, 125. Geotechnical Special Publications, ASCE/Geo-Institute16–49.
Lee, I.M., Lee, J.H., 1996. Prediction of pile bearing capacity using artificial
neural networks. Journal of Computer and Geotechnics 18 (3), 189–200.
Long, J.H., Kerrigan, J.A., Wysockey, M.H. 1999. Measured Time Effects for
Axial Capacity of Driven Piling. Transportation Research Record 1663,
Paper No.99-1183, pp. 8–15.
Maier, H.R., Dandy, G.C., 2000. Applications of artificial neural net-works to
forecasting of surface water quality variables: issues, applications and
challenges. In: Govindaraju, R.S., Rao, A.R. (Eds.), Artificial Neural
Networks in Hydrology. Kluwer, Dordrecht, The Netherlands, pp.
287–309.
Masters, T., 1993. Practical Neural Network Recipes in C++. Academic, San
Diego.
Ng, K., Roling, M., AbdelSalam, S., Suleiman, M., Sritharan, S., 2013a. Pile
setup in cohesive soil. I: experimental investigation. Journal of Geotechni-
cal and Geoenvironmental Engineering – ASCE 139 (2), 199–209.
 B. Tarawneh / Soils and Foundations 53 (2013) 607–615
Ng, K., Suleiman, M., Sritharan, S., 2013b. Pile setup in cohesive soil. II:
analytical quantifications and design recommendations. Journal of Geo-
technical and Geoenvironmental Engineering – ASCE 139 (2), 210–222.
Najjar, Y.,Basheer, I. 1996. A neural network approach for site characterization
and uncertainty prediction. ASCE Geotechnical Special Publication, vol.
58(1), pp. 134–148.
Nejad, F.P., Jaksa, M.B., 2011. Prediction of pile settlement using artificial
neural networks based on standard penetration test data. Journal of
Computer and Geotechnics 36 (7), 1125–1133.
Ornek, M., Laman, M., Demir, A., Yildiz, A., 2012. Prediction of bearing
capacity of circular footings on soft clay stabilized with granular soil. Soils
and Foundations 52 (1), 69–80.
Principe, J., Euliano, N., Lefebvre, W., 1999. Neural and Adaptive Systems:
Fundamentals through Simulations. Wiley, New York.
Rumelhart, D.E, Hinton, G.E, Williams, R.J., 1986. Learning internal
representation by error propagation. In: Rumelhart, D.E., McClelland, J.
L. (Eds.), Parallel Distributed Processing, vol. 1. MIT Press, Cambridge,
MA (Chapter 8).
Shahin, M. 2003. Use of Artificial Neural Networks for Predicting Settlement
of Shallow Foundations on Cohesionless Soil, University of Adelaide.
Shahin, M.A., Maier, H.R., Jaksa, M.B., 2002. Predicting settlements of
shallow foundations using artificial neural networks. Journal of Geotech-
nical and Geoenvironmental Engineering – ASCE 128 (9), 785–793.
Shahin, M.A., Maier, H.R., Jaksa, M.B., 2004. Data division for developing
neural networks applied to geotechnical engineering. Journal of Geotechnical
and Geoenvironmental Engineering-ASCE 18 (2), 105–114.
Skov, R. Denver, H. 1988. Time-dependence of bearing capacity of piles. In:
The 3rd International Conference on Application of Stress-wave Theory to
Piles, Ottawa, Canada, pp. 879–888.
615
Smith, M., 1993. Neural Networks for Statistical Modeling. Van Nostrand-
Reinhold, New York.
Stone, M., 1974. Cross-validatory choice and assessment of statistical
predictions. Journal of Royal Statistical Society B 36, 111–147.
Svinkin, M.R., 1996. Setup and relaxation in glacial sand-discussion. Journal
of Geotechnical and Geoenvironmental Engineering – ASCE 122 (4),
319–321.
Svinkin, M.R., Skov, R. 2000. Set-up effect of cohesive soils in pile capacity.
In: The 6th International Conference on Application of Stress-wave Theory
to Piles. Sao Paulo, Brazil, pp. 107–111.
Thompson, W.R., Held, L., Say, S., 2009. Test pile program to determine axial
capacity and pile setup for the Biloxi Bay Bridge. Deep Foundation
Institute 3 (1), 13–22.
Transportation Research Board 1999. Use of Artificial Neural Networks in
Geomechanical and Pavement Systems, Transportation Research Circular
E-C012, Prepared by A2K05(3) Subcommittee on Neural Nets and Other
Computational Intelligence-based Modeling Systems, Transportation
Research Board, Washington, USA.
Tokar, S.A., Johnson, P.A., 1999. Rainfall-runoff modeling using artificial
neural networks. Journal of Hydrologic Engineering – ASCE 4 (3),
232–239.
Twomey, J. Smith A. 1997. In: N. Kartam, I. Flood, J.H. Garrett (Eds.),
Artificial Neural Networks for Civil Engineers: Fundamentals and Applica-
tions, ASCE, New York, pp. 44–64.
Yan, W.M., Yuen, K.V., 2010. Prediction of pile set-up in clays and sands.
IOP Conference Series: Materials Science and Engineering 10, 1–8.
Zeghal, M., Khogali, W., 2005. Predicting the Resilient Modulus of Unbound
Granular Materials by Neural Networks. BCRA, Trondheim, Norway1–9.